From b793811ef6689cb701b0146196f47588c68f8bf0 Mon Sep 17 00:00:00 2001
From: Jason Tang <jason.tang@amd.com>
Date: Sun, 20 Jun 2021 14:24:24 -0400
Subject: [PATCH] SWDEV-287088 - Workaround VM page fault on Windows

Some chunk memory are not guaranteed to be resident during
initial allocation. Use CPDMA to force resident.

Change-Id: If1a2da3e75f136caaa4c7a29d8f604d6af2639fa
---
 rocclr/device/pal/palresource.cpp | 24 ++++++++++++++++++++++++
 rocclr/device/pal/palresource.hpp |  1 +
 2 files changed, 25 insertions(+)

diff --git a/rocclr/device/pal/palresource.cpp b/rocclr/device/pal/palresource.cpp
index 2950e8a01f..99e8d1c344 100644
--- a/rocclr/device/pal/palresource.cpp
+++ b/rocclr/device/pal/palresource.cpp
@@ -1929,6 +1929,22 @@ bool MemorySubAllocator::InitAllocator(GpuMemoryReference* mem_ref) {
   return true;
 }
 
+// ================================================================================================
+void MemorySubAllocator::forceResident(GpuMemoryReference* mem_ref) {
+  if (IS_WINDOWS) {
+    // Write one DWORD using CPDMA to force resident
+    GpuEvent event;
+    auto gpu = device_->xferQueue();
+    uint32_t data = 0;
+
+    gpu->eventBegin(MainEngine);
+    gpu->queue(MainEngine).addCmdMemRef(mem_ref);
+    gpu->iCmd()->CmdUpdateMemory(*mem_ref->iMem(), 0, 4, &data);
+    gpu->eventEnd(MainEngine, event);
+    gpu->waitForEvent(&event);
+  }
+}
+
 // ================================================================================================
 bool MemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved_va) {
   Pal::GpuMemoryCreateInfo createInfo = {};
@@ -1944,6 +1960,8 @@ bool MemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved_va) {
   createInfo.mallPolicy = static_cast<Pal::GpuMemMallPolicy>(device_->settings().mallPolicy_);
   GpuMemoryReference* mem_ref = GpuMemoryReference::Create(*device_, createInfo);
   if (mem_ref != nullptr) {
+    // Workaround: some chunk memory are not guaranteed to be resident during initial allocation.
+    forceResident(mem_ref);
     return InitAllocator(mem_ref);
   }
   return false;
@@ -1964,6 +1982,8 @@ bool CoarseMemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved_va) {
   createInfo.mallPolicy = static_cast<Pal::GpuMemMallPolicy>(device_->settings().mallPolicy_);
   GpuMemoryReference* mem_ref = GpuMemoryReference::Create(*device_, createInfo);
   if (mem_ref != nullptr) {
+    // Workaround: some chunk memory are not guaranteed to be resident during initial allocation.
+    forceResident(mem_ref);
     return InitAllocator(mem_ref);
   }
   return false;
@@ -1980,6 +2000,8 @@ bool FineMemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved_va) {
   createInfo.mallPolicy = Pal::GpuMemMallPolicy::Never;
   GpuMemoryReference* mem_ref = GpuMemoryReference::Create(*device_, createInfo);
   if ((mem_ref != nullptr) && InitAllocator(mem_ref)) {
+    // Workaround: some chunk memory are not guaranteed to be resident during initial allocation.
+    forceResident(mem_ref);
     mem_ref->iMem()->Map(&mem_ref->cpuAddress_);
     return mem_ref->cpuAddress_ != nullptr;
   }
@@ -1998,6 +2020,8 @@ bool FineUncachedMemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved
   createInfo.mallPolicy = Pal::GpuMemMallPolicy::Never;
   GpuMemoryReference* mem_ref = GpuMemoryReference::Create(*device_, createInfo);
   if ((mem_ref != nullptr) && InitAllocator(mem_ref)) {
+    // Workaround: some chunk memory are not guaranteed to be resident during initial allocation.
+    forceResident(mem_ref);
     mem_ref->iMem()->Map(&mem_ref->cpuAddress_);
     return mem_ref->cpuAddress_ != nullptr;
   }
diff --git a/rocclr/device/pal/palresource.hpp b/rocclr/device/pal/palresource.hpp
index ec11a1a24b..bf3cf0196b 100644
--- a/rocclr/device/pal/palresource.hpp
+++ b/rocclr/device/pal/palresource.hpp
@@ -528,6 +528,7 @@ class MemorySubAllocator : public amd::HeapObject {
   //! Allocate new chunk of memory
   virtual bool CreateChunk(const Pal::IGpuMemory* reserved_va);
   bool InitAllocator(GpuMemoryReference* mem_ref);
+  void forceResident(GpuMemoryReference* mem_ref);
 
   Device* device_;
   std::unordered_map<GpuMemoryReference*, MemBuddyAllocator*> heaps_;